Background

Causal inference [1,2,3] is an important tool in the domain of health sciences for informatics work such as finding causal effects of an adverse outcome or risk factors for a disease. Causality has traditionally been a core concept across all branches of medical science and considered when diagnosing patients based on their symptoms, effects of treatment, and years of historical evidence [4]. Indeed, the study of causal inference in health science research dates back to the 1970s and 1980s [5, 6]. Yet, non-causal models such as regression tend to be more commonly applied in health science research than causal models [7]. The goals and philosophy of causal inference differ from those of association-based predictions in several ways. For instance, with predictive models such as regression, one wants to measure the likelihood of occurrence of an event as a result of another event; for example, the occurrence of lung cancer based on exposure to smoke in the environment. However, such predictions may be subject to confounding; for instance, a researcher may find that a regional increase in the sale of matches is associated with lung cancer, but not necessarily causally associated if, for example, the association actually reflects a regional increase in match sales due to frequent blackouts, with a secondary unrelated association related to lung cancer and perhaps attributable to exposures such as workplace chemical exposures or lack of healthcare access. Most predictive models, unlike causal inference models, do not readily account for confounding variables and hence cannot differentiate causal versus spurious associations. Another aspect of causal inference that differentiates it from non-causal models is the ability to provide an explanation for the relationship between two events. For instance, causal inference can help to discern why a patient is sick and diagnose them or identify medications to treat them based on the underlying cause of their symptoms.

Electronic health records (EHRs) provide a potential source of structured clinical data such as diagnoses, medications, and laboratory results. Access to EHR data is thus critical for the advancement of health science research and clinical research. However, the many federal and institutional regulations that surround clinical data, while necessary to ensure patient privacy and protection of sensitive data, limit access to the data for research. In this study, we analyzed a patient-level dataset extracted from a regulatory-compliant open service called the Integrated Clinical and Environmental Exposures Service (ICEES) [8]. ICEES supports several use cases, including asthma, drug-induced liver injury, and coronavirus infection. The ICEES data are constructed by integrating clinical data elements derived from patient EHRs and environmental exposures data derived from a variety of public sources of environmental exposures data [9]. The data are then binned and de-identified by stripping all protected health information per the Safe Harbor method of the Health Insurance Portability and Accountability Act. The ICEES data are then exposed via an open application programming interface (OpenAPI). For our principal application use case, we focus on an existing ICEES cohort of patients with asthma or a related common pulmonary disorder (see [8] for details). We asked if there is a causal relationship between asthma attacks and the following features: sex, race, prednisone use, diagnosis of obesity, residential proximity to a major roadway or highway, residential density, and exposure to airborne pollutants. We focus on these features because published studies, including our prior work [8,9,10], have recognized one or more of them to be associated with asthma attacks. We consider the number of annual emergency department (ED) or inpatient hospital visits for respiratory issues as the primary outcome measure and indicator of asthma attacks, as we have done previously. We first generate a causal inference network. We then demonstrate simulated external interventions as an approach to validate the inferred causal network. We use subject matter expert knowledge and publication support as our ground truth to measure the correctness of our causal inference model. Finally, we discuss our findings, including the benefits and limitations of our causal inference model and approach.

Methods

Generation of multivariate ICEES table

We focused on an existing ICEES cohort of patients with asthma or another common pulmonary disorder and examined outcomes over a one-year study period (see [8] for details on the inclusion and exclusion criteria). In brief, the patients were included if they had at least one diagnosis of asthma and/or another common respiratory disorder, had a prescription or administration of a drug typically used to treat asthma and/or other common respiratory disorders, or had frequent ED visits during which an albuterol nebulizer was administered. The majority of patients included in the final dataset were 45 years of age or older, female, non-Hispanic white, and residing in a rural region.

We asked if there is a causal relationship between asthma attacks and the following features: sex, race, prescriptions for prednisone use, diagnosis of obesity, residential proximity to a major roadway or highway, residential density, and exposure to high levels of airborne pollutants. We selected these features because published studies, including our prior work [8,9,10], have recognized one or more of them to be associated with asthma attacks. The racial categories were self-reported, as defined according to our hospital’s EHR system, and included to capture potential racial disparities for further investigation into whether those are related to socioeconomic conditions, healthcare access, or other factors. We defined “asthma attacks” based on the annual number of ED or inpatient visits for respiratory issues. This is an acceptable clinical proxy for asthma exacerbations, one that we have applied successfully in prior work [8,9,10].

We queried the ICEES OpenAPI to generate a multivariate table. We focused on eight ICEES feature variables, namely, TotalEDInpatientVisits, Sex, Race, Prednisone, Obesity, PM2.5 Exposure, RoadwayExposure, and EstResidentialDensity, and TotalEDInpatientVisits, as defined in Table 1. The majority of patients did not have any ED or inpatient hospital visits over the one-year study period and were not active in the year of interest (data not shown), meaning that their EHR did not indicate any healthcare utilization. This finding was expected, but it introduced a skew in the distribution of TotalEDInpatientVisits, with the vast majority of patients grouped as TotalEDInpatientVisits = 0. To minimize the skew, we applied the filter “Active_In Year” before extracting the multivariate table, with Active_In_Year = 1 to select only patients who were active in 2010. In Fig. 1, we show the distribution of TotalEDInpatientVisits among the discrete categories of each feature variable after applying the Active_In_Year filter. With the Active_In_year filter in place, the distribution of TotalEDInpatientVisits indicated that most patients who were active in year 2010 visited the ED or had an inpatient hospital visit at least one time over the year of interest. There was an imbalance in TotalEDInpatientVisits across levels for some feature variables such as Prednisone, Obesity, Race, RoadwayExposure, and PM2.5Exposure. The final multivariate table in this work comprised data on 14,937 patients (i.e., rows represented individual patients in the asthma cohort, and columns represented feature variables). Figure 1 shows the number of TotalEDInpatientVisits across each level of the feature variables.

Table 1 Feature variables used to generate the multivariate table
Full size table
Fig. 1
figure 1

Stacked bar chart representing the number of TotalEDInpatientVisits across each level of the feature variables. See Table 1 for feature variable definitions

Full size image

Evaluation of feature importance

We evaluated the importance of each feature using a tree-based machine learning model: random forest. The random forest analysis was conducted to provide a comparison with the causal network analysis. We leveraged the caret R package [11] to evaluate the feature importance. We controlled the parameters for training by using the repeatedcv method to divide our dataset into ten-folds cross-validation and repeated three times.

Causal network analysis

Most of the naturally occurring trends that we come across are simply passive observations of events occurring in the world that are either coincidental or unexplained associations. For example, statements like “drinking beer everyday increases the chance of prostate cancer” are common in the news and scientific reporting and in our day-to-day personal beliefs. These associations can be easily mistaken as causation, making us susceptible to logical fallacies without knowing the real underlying cause. Causal inference is the science of distinguishing cause from effect [1,2,3]. It is an important field of research because it helps us eradicate spurious correlation [12,13,14]. The primary aim of inferring causal relations from data is to discover interactions between different entities in the form of Vi → Vj , where Vi and Vj are observable features in a domain and the arrow indicates that the state of Vi influences the state of Vj. Causal inference can be either discovered through observational measurements (seeing) or from measurements after performing some external manipulation/intervention (doing). A causal network [1,2,3, 15] can be represented with a directed acyclic graph (DAG) G = (V, E), where V = Vi, . . . .., Vn denotes the set of features and E ∈ (V × V ) denotes the set of edges that are causal in nature. For a causal edge (Vi, Vj ), we say that Vi is a cause (parent) of Vj , and Vj is the resulting effect (child) of Vi. Let pa(Vi) denote the set of parents of Vi. The conditional probability distribution Pi defines the probability of Vi given the state of its parents pa(Vi). A causal network represents a joint distribution P over variables V as long as it satisfies two main assumptions:

  • (a) Causal Markov assumption: Any given variable Vi is independent of its non-descendants, conditioned on all of its direct causes (parents). This implies that the joint distribution P(V) can be factored as:

  • p(V) = ∏ni=1 pi (Vi | Pa(Vi)).

  • (b) Faithfulness assumption: The joint distribution p(V1,..., Vn) is faithful to G if every conditional independence relation in the probability distribution P is entailed by the Markov assumption applied to G [16].

To reconstruct a causal graph from data, we generally start by finding an approximation of the graph, given V, and then optimize based on conditions on data. The two main approaches used for causal network inference are:

  1. 1.

    Score-based: This is based on a Bayesian scoring function S(G | D), which estimates the goodness-of-fit of graph G to the data D [17], as objective functions to maximize, while favoring simpler structures. The score function is usually combined with a search heuristic that explores the space of all possible graphs. Score-based methods are robust and can be extended to include interventional studies (if available), but they are not scalable as network or data size increases.

  2. 2.

    Constraint-based: This method is based on estimating some of the conditional (in)dependencies in the distribution P from the data D by performing hypothesis tests of conditional independence. Constraint based methods usually start with a fully connected, undirected graph and progressively remove edges whenever a new conditional independence relation is discovered, while satisfying the corresponding d separation statements.

In this work, we used a constraint-based approach called the Principal Component (PC) algorithm, given that the dataset was observational. To infer the causal graph from data, we learned the equivalence class of a directed acyclic graph (DAG) from data with the traditional constraint-based PC algorithm proposed by [15]. Given a dataset D having n features Vi,....., Vn, we conducted the following steps. We started with a complete undirected graph given n features. We then eliminated edges between variables that are unconditionally independent. For each pair of variables (Vi, Vj) with an edge between them, and for each variable Vk with an edge connected to either of them, we eliminated the edge between Vi and Vj if Vi ⊥⊥ Vj | Vk. For each pair of variables Vi, Vj having an edge between them, and for each pair of variables Vk, Vl with edges both connected to Vi or both connected to Vj, we eliminated the edge between Vi and Vj if Vi ⊥⊥ Vj | Vk, Vl. We continued to check independencies conditional on subsets of variables of increasing size n until there were no more adjacent pairs (Vi, Vj) such that there was a subset of variables of size n in which all of the variables in the subset were adjacent to Vi or adjacent to Vj. For each triple of variables (Vi, Vj, Vk) such that Vi and Vj were adjacent, Vj and Vk were adjacent, and Vi and Vk were not adjacent, we oriented the edges Vi––Vj––Vk as Vi → Vj ← Vk, if Vj was not in the set conditioning on which Vi and Vk became independent and the edge between them was accordingly eliminated. We called such a triple of variables a v-structure. For each triple of variables such that Vi → Vj––Vk, and Vj and Vk were not adjacent, we oriented the edge Vj––Vk as Vj → Vk (i.e., orientation propagation).

We applied a causal model based on the eight feature variables included in our random forest analysis (section “Evaluation of feature importance”). We compared our model output with a model of expected edges based on subject matter expertise (e.g., a distinguished professor, practicing physician, and expert on pulmonary disorders) and the published literature [18,19,20,21,22,23,24,25,26]. Thus, both sources were used to generate a model of expected edges.

2.4 Simulated interventions.

We used the eight-feature causal model generated as described in section “Causal network analysis” to answer relevant questions through inference. To evaluate this, we computed the effects of interventions on features by modifying the network to simulate interventions. First, we removed undirected edges. We then learned the parameters of our learned causal DAG, given the network structure and the data. Next, we constructed a mutilated network to simulate a perfect intervention by setting a target node to a particular value. Finally, we tested the effects of three interventions on TotalEDInmpatientVisits: Obesity = 1 (all patients forced to be obese); Prednisone = 1 (all patients forced to be using prednisone); and Sex = Male (all patients forced to be male). The expectations, based on the causal inference network developed under section “Causal network analysis”, were that interventions on obesity and prednisone would have direct effects on the number of TotalEDInpatientVisits, whereas an intervention on sex would not have direct effects. We note that while the interventions on obesity and prednisone are feasible, we recognize that an intervention on sex is not; however, we included sex as a test of the causal model and our assumptions, not its realistic implementation.

Results

Feature importance

In our feature importance analysis using a random forest model, we found that Prednisone, Race, Obesity, RoadwayExposure, and PM2.5Exposure were the main contributing factors to asthma attacks (Fig. 2).

Fig. 2
figure 2

Relative feature importance for all features with respect to TotalEDInpatientVisits. See Table 1 for feature variable definitions

Full size image

Causal analysis

Having completed the random forest analysis, we then conducted an independent causal analysis. First, we applied a PC algorithm to the ICEES multivariate feature table using the same eight feature variables used for the random forest analysis. In Fig. 3, we show the inferred casual graph. Expected relationships between features based on subject matter expertise and published literature are represented in black lines (solid and dashed, respectively). There were eight such expected edges, which we used to measure the structure learning accuracy of the causal algorithm. Solid black lines represent expected edges (true positives) that were reported via the PC algorithm, while dashed lines represent edges that were expected but missed (false negatives). Newly found relationships inferred by the PC algorithm, that were not expected, are represented in red (false positive). We note that there were a few undirected edges detected, for which the algorithm was not able to determine directionality.

Fig. 3
figure 3

Inferred causal graph. Solid black lines represent inferred expected edges based on subject matter expertise combined with published literature (true positives), dashed lines represent missed expected edges (false negatives), and red lines represent unexpected edges, meaning not expected based on subject matter expertise or the published literature (false positives)

Full size image

Three of eight expected edges as determined by subject matter expertise were inferred; two out of three additional edges expected edges as reported in the literature were inferred (see section “Causal network analysis” for details). The expected directed edge from Race → TotalEDInpatientVisits was missed.

Effects of Intervention

Having learned a causal network from the data, we then used it to answer relevant questions by making inferences. To evaluate the network, we tested the effects of three simulated interventions on TotalEDInpatientVisits. Specifically, to substantiate the causal relationships identified section “Causal analysis”, we tested the effects of interventions based on the following expected claims:

  • Claim (a). Obesity should have a direct effect on TotalEDInpatientVisits. Hence, conducting an intervention on the node “Obesity” (i.e., forcing all patients to be obese) should produce a direct change (increase or decrease, accordingly) in the probability distribution of TotalEDInpatientVisits.

  • Claim (b). Prednisone should have a direct effect on TotalEDInpatientVisits. Hence, conducting an intervention on the node “Prednisone” (i.e., focusing all patients to be using prednisone) should produce a direct change (increase or decrease, accordingly) in the probability distribution of TotalEDInpatientVisits.

  • Claim (c). Sex should not have a direct effect on TotalEDInpatientVisits, as our causal network identified only an indirect effect of sex by way of a direct effect on obesity. Hence, conducting an intervention on the node “Sex” (i.e., forcing all patients to be male) should not produce a direct change (increase or decrease, accordingly) in the probability distribution of TotalEDInpatientVisits.

We conducted these three simulated interventions on our learned causal network. To test Claim (a), we created a mutilated network by fixing the state of Obesity to 1, which means we forced Obesity to be present. For Claim (b), we fixed the state of Prednisone to be 1, meaning that we forced prednisone use to be present. For Claim (c), we fixed the state of Sex to be Male. Next, we compared the changes in the probability distribution of TotalEDInpatientVisits before and after these three ad hoc simulated interventions to confirm the expected causal influences (Fig. 4). The change in the probability distribution for TotalEDInpatientVisits for interventions (a) and (b) shifted to the right with each intervention due to their causal relationships to the outcome: 0.5681 to 0.6642 mean number of visits (9.62% increase) for obesity (Fig. 4a); 0.5681 to 0.7271 mean number of visits (15.90% increase) for prednisone (Fig. 4b). For intervention (c), the change in the probability distribution before and after the intervention was negligible (Fig. 4c): 0.5681 to 0.5722 mean number of visits (0.42% increase) for sex.Thus, intervening on obesity and prednisone caused a shift to the right in the number of annual ED or inpatient visits for respiratory diseases, as expected, given that our causal model showed direct effects of each variable on the outcome. In contrast, intervening on Sex had a negligible effect on the probability distribution of TotalEDInpatientVisits, also as expected, given that our causal model showed only an indirect effect of sex on the outcome by way of obesity.

Fig. 4
figure 4

The change in the mean number (% increase) of TotalEDInpatientVisits after each intervention: a 0.5681 to 0.6642 mean number of visits (9.62% increase) for Obesity; b 0.5681 to 0.7271 mean number of visits (15.90% increase) for Prednisone; and c 0.5681 to 0.5722 mean number of visits (0.42% increase) for Sex. Interv = intervention

Full size image

Discussion

In this paper, we demonstrated the ability to use the ICEES OpenAPI to answer important questions about causal relationships between factors affecting asthma attacks. We focused on a large cohort of patients with asthma or related conditions and a dataset that included data derived from EHRs and a variety of public sources of environmental exposures data. We selected eight feature variables for our analyses; namely, sex, race, obesity, prednisone use, airborne particulate matter exposure, major roadway/highway exposure, residential density, and annual number of ED or inpatient hospital visits for respiratory issues. The racial categories were self-reported, as defined according to our hospital’s EHR system, and included to capture potential racial disparities for further investigation. We applied a random forest algorithm and identified prednisone, race, and obesity as significant predictors of annual ED or inpatient visits for respiratory issues, followed by residential distance from a major roadway/highway, airborne particulate exposure, and sex. We then applied an independent causal inference model to the data, using the same feature variables, and found that prednisone and obesity were causally related to annual ED or inpatient visits, and sex and race were found to be indirectly related to annual ED or inpatient visits via a causal relationship to obesity. To validate our causal model, we then performed simulated interventions based on our causal network. Specifically, we tested the effects of “forcing” all patients to be obese, using prednisone, and of the male sex. As expected, we found that forcing all patients to be obese or using prednisone had a direct effect on annual ED or inpatient visits, whereas forcing all patients to be male did not have a direct effect. The results of our interventions, while carrying an undefined degree of statistical uncertainty, generally support our causal network analysis. Indeed, one of the strengths of causal analysis modeling, unlike predictive modeling, is that it minimizes the influence of confounding. Nonetheless, confounding remains a consideration due to factors that were unaccounted for such as physician prescribing practices regarding the use of prednisone.

Our results are largely consistent with previously published literature. For instance, prednisone, which is commonly prescribed for patients who are non-responsive to first-line treatments such as inhaled albuterol [18], has been identified as a factor associated with asthma exacerbations and ED or inpatient visits for respiratory issues [19]. Female sex, obesity, and Black African American race have previously been identified as factors that contribute to asthma attacks [20]. In another work by our group [10] and others [21], obesity and sex were found to be highly related to asthma attacks. Several other works [9, 22] have additionally found a significant association between Black African American race and increased risk of asthma attacks. Exposure to major roadways or highways has also been found to be a risk factor for asthma. Several studies [23, 24] have demonstrated an increase in asthma attacks among patients residing in close proximity to a major roadway or highway. Our findings on the relationship between roadway exposures and asthma exacerbations have been inconsistent, with evidence to support [20] and negate [25] a relationship.

One factor that we expected to find in our model as causally related to asthma attacks, but did not, is exposure to airborne particulate matter. Exposure to airborne particulate matter is a well-established trigger for asthma attacks [8, 9, 19, 20, 25, 26]. The failure to detect a causal relationship between exposure to airborne particulate matter and asthma attacks likely reflects the imbalance in the distribution of patients across bins. Indeed, we are actively refining both our exposure models and our binning strategy. For instance, instead of using a Python algorithm to bin the airborne pollutant exposures, we are considering a binning strategy based on subject matter expertise alone.

Conclusions

EHR data, while being a rich data source for important clinical information, are mostly observational and generally challenging to access due to regulatory constraints. Performing real-world interventions are not only costly, but even impractical, given the need to integrate large data sources across various domains. Causal inference provides an excellent tool to simulate clinical interventions and answer questions about the effects of medical and healthcare interventions. In this study, we used the regulatory-compliant open ICEES service to generate a multivariate feature table and apply a causal inference model, as well as conduct simulated interventions, to explore the influence of key demographic factors and environmental exposures on asthma attacks. Our results were largely consistent with expectations based on subject matter expert opinion and the published literature. As part of our future studies, we are expanding our causal inference model to include additional features and additional years of data in order to reflect the underlying causal relationships at a larger scale, while supporting additional use cases, including a cohort of patients with primarily ciliary dyskinesia or another rare respiratory disorder.